Systems in various states of evolution.

The problem

Finding good deals for Canada's Chinese-speaking community meant manually searching fragmented sources, rewriting product descriptions, and posting across multiple platforms by hand — every single day.

What I built

Built the platform end-to-end: FastAPI + Next.js 14 + MongoDB + Meilisearch, containerized with Docker Compose. An automated Playwright crawler bypasses Amazon bot-detection via Chrome DevTools Protocol. A pluggable AI-provider layer (OpenAI / Claude / Deepseek, hot-swappable via one env variable) auto-generates Traditional Chinese titles, categories, and feature summaries. Facebook Graph, Instagram Stories, and Threads APIs handle scheduled auto-posting.

Manual work it replaces

Manual deal hunting, manual description rewriting, manual social posting — the entire content pipeline now runs on its own.

What I learned

Proactive security work matters: I audited the system myself and patched 3 critical vulnerabilities including hardcoded DB credentials, an X-Forwarded-For rate-limit bypass, and an XSS-readable admin key. The pluggable AI layer taught me that swapping cost/quality trade-offs should never require touching business logic.

Built with

Status

Live at canadeal.ca.

The problem

Grocery pricing has almost no transparency. ~12,500 T&T products across 2 stores and 3 languages with no way to track how prices actually move over time.

What I built

A Node.js crawler on a daily crontab + Docker Compose schedule. Two-phase crawl pipeline (bulk category listing, then detail fetch only for items with options) reduces unnecessary requests by ~95%. Polite rate-limiting with p-limit, random jitter, and exponential back-off keeps full runs under 5 minutes per combination. MongoDB price-history sub-documents append only on price change, cutting write volume by ~80%. Replaced Elasticsearch with Meilisearch to eliminate hardcoded-credential risk while keeping multi-language full-text search.

Manual work it replaces

Manual price-checking and the guesswork of "is this actually a good price?" — replaced with a history that accumulates on its own, every day.

What I learned

A race condition was serving empty API responses because of a delete-then-insert pattern; fixed with idempotent bulk upserts. Even a fairly simple system becomes powerful once data accumulates continuously — the value isn't in any single reading, it's in the history.

Built with

Status

Live at tnt.canadeal.ca.

The problem

Sharing affiliate links at scale means managing custom aliases, tracking clicks, injecting affiliate tags consistently, and splitting traffic for A/B tests — none of which basic link shorteners support.

What I built

Full-stack FastAPI + MongoDB platform with custom alias support, QR code generation, and deep link routing. A/B testing for redirect targets with configurable weight distribution, scoped by subscription plan tier (Free / Pro / Business). Affiliate tag injection with per-domain rules and an ethical no-overwrite policy — existing affiliate parameters are never replaced. Click analytics pipeline tracking device type, geo-location, referrer, and time-series data with EPC metrics via REST API. JWT authentication with role-based access control.

Manual work it replaces

Manual link management, manual affiliate tag insertion, and guesswork about which redirect targets actually convert.

What I learned

The ethical no-overwrite policy on affiliate tags was a deliberate design decision, not an afterthought. Building plan-gated feature enforcement consistently across every endpoint is the kind of unglamorous work that makes a real product feel solid.

Built with

Status

Live at go.canadeal.ca.

The problem

Flashfood has real discounted-grocery opportunities, but discovery is fragmented across cities, stores, and categories — with no good way to filter or track what's available.

What I built

Live deal feed using Firestore onSnapshot (no polling) with composable region/store/category/keyword filters and multiple sort modes. Infinite scroll via IntersectionObserver in 32-item batches. Per-city filter persistence in sessionStorage. Multi-language UI (EN / ZH / FR / KR / JP). Dynamic Open Graph images generated at the Edge for rich social-sharing previews.

Manual work it replaces

The manual hunt-and-sort of finding and tracking discounts across a noisy, constantly changing source.

What I learned

A filter-restore race condition on city switch required a ref-based guard to fix cleanly. Deriving context-aware category pills through a two-pass filtering strategy turned out to be the right call — it eliminated a whole class of inconsistency.

Built with

Status

Live at flashfood.canadeal.ca.

The problem

Tailoring a resume to a specific job description is repetitive, time-consuming, and easy to get wrong — especially when you have a large knowledge base of skills, projects, and experience to draw from.

What I built

Full-stack resume builder: Next.js 15 (App Router, TypeScript, TailwindCSS) frontend + FastAPI backend, deployed on Vercel + Docker. OpenAI API analyzes job descriptions and recommends the most relevant skills, projects, and experience entries from a structured knowledge base, with heuristic fallback for offline usage. Personal profile database with full CRUD for skills, work experience, and projects — supporting both manual entry and AI-assisted bulk import from raw text. Resume export in PDF, DOCX, Markdown, and plain text. A privacy-conscious AI rewrite feature polishes bullet points without fabricating facts, inflating responsibilities, or inventing metrics.

Manual work it replaces

The manual process of reading a job description, cross-referencing your own experience, and deciding what to include — replaced with a system that does the matching and drafting automatically.

What I learned

JWT authentication with per-user OpenAI API key management — masking stored keys on retrieval — is the kind of detail that separates a real product from a demo. The privacy constraint on the AI rewrite (no fabrication, no inflation) was a harder design problem than the AI integration itself.

Built with

Status

Live at resume.canadeal.ca.